finsql-mlx-qwen3-4b-instruct-4bit

Latest Finetuning

HF Model Path: gccmorgoth/finsql-mlx-qwen3-4b-instruct-4bit

Finetuning Details

• Method: Direct Preference Optimization (DPO)
• Checkpoint: Iteration 300
• Validation Loss: 0.048
• Training Loss: 0.122
• Learning Rate: Cosine decay with warmup
• LoRA Rank: 16

Performance

• Validation loss: 0.048 (optimal convergence point)
• Selected at iteration 300 to prevent overfitting
• DPO training for improved preference alignment on financial SQL tasks

Model Selection

• Checkpoint: Iteration 300 selected based on validation loss curve
• Rationale: Optimal balance between training convergence and generalization
• Training Dynamics: Early stopping before overfitting (val loss increased at iter 700+)

Dataset

Usage

Database: [database_name]

Task

Sample Prompt Format

Python